Meni (Menachem) Brief · Applied Scientist & Builder · Seattle, WA
Applied scientist and end-to-end builder. 3+ years at Microsoft training and evaluating LLMs, most recently RL for model behavior in complex tool use. Co-founded Simply-Useful and shipped a production mobile app as founding engineer and technical lead. Current work: proving an agent can be trusted before it touches anything real.
email available on request · linkedin.com/in/meni-brief · scholar.google.com/citations?user=t-P-tjgAAAAJ · semanticscholar.org/author/Meni-Brief/2217252086
experience
run(agent, controlled_world)
Most teams deploying agents have never defined what the agent is actually allowed and able to do. It stays implicit until something breaks, and observability only reports it after it happened to a real customer. I'm building the step before: run the agent in a high-fidelity world you control, capture and replay every run deterministically, and judge it by the side effects it causes rather than what it says it did. Started as a proof of concept, now building a production-grade system for real-world use.
ship(mobile_app → app_stores)
Founding engineer of a production-grade, mobile-first productivity app: from zero to live app-store releases with real users, leading a small team I hired. I owned the system architecture end to end: Django backend on DigitalOcean, in-app AI processing, two-way sync with Google services, CI/CD on GitHub Actions, product analytics, and error monitoring, all on an AI-native development workflow that kept a micro team shipping at production velocity.
rl.improve(tool_use, mcp)
Applied RL to improve LLM behavior in complex tool-use scenarios: agent orchestration and correct invocation of MCP tools. Led applied research on domain adaptation and knowledge injection: data curation, experiment design, training, evaluation, and results communication to internal and external stakeholders. Delivered LLM success stories in healthcare and financial services; co-authored research papers and one patent.
research(llm.applications)
Original research on LLM applications and capabilities across industries (PyTorch / Transformers / OpenAI), from experiment design through implementation; improved ML models running in production.
segment(ultrasound) → fda_cleared
Built deep-learning segmentation models for ultrasound imaging; the resulting tool (CNerve) passed FDA requirements for AI in healthcare. Served as data owner and collaborated directly with physicians.
research
A systematic comparison of unsupervised fine-tuning and retrieval-augmented generation for injecting knowledge into LLMs, across knowledge-intensive tasks, including entirely new facts models never saw in pre-training.
skills
education
MSc in Data Science (GPA 91.4). BSc in Information Systems Engineering, cum laude, #3 in class (GPA 90.6), with minors in Economics and Fintech.